| --- |
| license: cc-by-nc-sa-4.0 |
| task_categories: |
| - text-to-speech |
| - audio-classification |
| language: |
| - en |
| tags: |
| - deepfake-audio |
| - synthetic-speech |
| - voice-cloning |
| - deepfake-detection |
| --- |
| |
| # 4TTSDeepfakeDataset |
|
|
| ## Description |
| Based on the [Hi-Fi Multi-Speaker English TTS Dataset](https://www.openslr.org/109/), this dataset includes 148.06 hours and 75,575 utterances of 24kHz Mono 16-bit audio. It uses the utterances of speakers 6097, 6670, 6671, and 9017 from the original corpus. |
|
|
| For each original bona fide utterance, there are 4 corresponding deepfake zero-shot 1:1 utterances generated by the following models (15,115 utterances each): |
| * [F5-TTSv1](https://huggingface.co/SWivid/F5-TTS) |
| * [MaskGCT](https://huggingface.co/amphion/MaskGCT) |
| * [CosyVoice2](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B) |
| * [Fish-S1mini](https://huggingface.co/fishaudio/s1-mini) |
|
|
| ## Dataset Structure |
| The dataset includes 5 folders with utterances and a `parallel_manifest.json` file that provides the 1:4 (bona fide to TTS deepfakes) mapping. |
|
|
| ### Recordings per TTS System/Speaker |
|
|
| | Speaker ID | Parallel Utterances | Bona Fide (hrs) | MaskGCT (hrs) | F5-TTSv1 (hrs) | Fish-S1mini (hrs) | CosyVoice2 (hrs) | |
| | :--------------- | :-----------------: | :-------------: | :-----------: | :------------: | :---------------: | :--------------: | |
| | **6097** | 3,500 | 6.98 | 7.00 | 6.85 | 6.83 | 7.11 | |
| | **6670** | 3,556 | 7.18 | 7.19 | 7.07 | 6.97 | 7.56 | |
| | **6671** | 4,538 | 9.02 | 8.93 | 8.74 | 8.57 | 9.11 | |
| | **9017** | 3,521 | 6.63 | 6.66 | 6.55 | 6.38 | 6.73 | |
| | **Total Corpus** | **15,115** | **29.82** | **29.77** | **29.20** | **28.76** | **30.51** | |
|
|
| ## Acknowledgements and Attribution |
|
|
| ### Original Audio Data |
| This dataset is a derivative work based on the [Hi-Fi Multi-Speaker English TTS Dataset](https://www.openslr.org/109/) created by E. Bakhturina, V. Lavrukhin, B. Ginsburg, and Y. Zhang. The original dataset is licensed under a [Creative Commons Attribution 4.0 International License (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/). |
|
|
| **Modifications to the Original Material:** 1. The original bona fide audio files from speakers 6097, 6670, 6671, and 9017 were modified (sliced and merged) to create the parallel baseline utterances. |
| 2. These preprocessed bona fide files were then used as the source material to generate 1:1 synthetic speech (deepfakes). |
|
|
| ### Synthetic Data Generation (Deepfakes) |
| The synthetic deepfake utterances in this dataset were generated using the following open-source Text-to-Speech models. The generated audio inherits the licensing terms of their respective source models: |
|
|
| * **[F5-TTSv1](https://huggingface.co/SWivid/F5-TTS)** by SWivid (Licensed under CC BY-NC 4.0) |
| * **[MaskGCT](https://huggingface.co/amphion/MaskGCT)** by Amphion (Licensed under CC BY-NC 4.0) |
| * **[Fish-S1mini](https://huggingface.co/fishaudio/s1-mini)** by Fish Audio (Licensed under CC BY-NC-SA 4.0) |
| * **[CosyVoice2](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B)** by FunAudioLLM (Licensed under Apache 2.0) |
|
|
| ## Usage Restrictions |
| In accordance with the licenses of the underlying datasets and TTS models (specifically the CC-BY-NC and CC-BY-NC-SA clauses), the valid usage of this dataset only covers **educational and personal purposes**. |
| Misuse of the dataset is your own responsibility. |
